Introducing VERSUR Loops

One pass from an Agent is often a first draft. Design work rarely ships that way. You want the block to keep going - draft, check the work, fix what is weak - until the result clears the standard you care about. Without turning the canvas into a handmade generate / evaluate / improve chain every time.

Today we are introducing VERSUR Loops on the Agent block - outer iteration over stacked procedures until a stop condition or max passes. Same block. Several passes. Each pass builds on the work so far, so the agent refines instead of restarting from zero.

Why loops live on the Agent

When quality is about refining one artifact - a brief interpretation, a zoning memo, a concept write-up, a visualization prompt - you do not need extra blocks on the canvas. A loop on the Agent keeps the method on the block: procedures for start, generate, evaluate, improve, research, plan, act, and more, stacked and repeated until the work is good enough.

That is different from a subflow loop playbook, where each pass spans multiple blocks (image generation plus scoring, multi-agent critique, tool-heavy retry graphs). Use an Agent loop for single-block refinement. Use a subflow when the iteration graph itself needs more than one block.

Built-in loops

Open the Loop field on an Agent, choose Add loop, and pick a built-in - or create your own. Four starters cover the common studio patterns:

• Creative refinement - Generate, evaluate against a rubric, improve. Repeat until quality or max passes. Strong for writing and design briefs.
• Architecture satisfaction - Plan, act, evaluate architecture quality, improve. Tool-heavy; one procedure per pass so act phases get full tool turns.
• Research then generate - Gather context, produce output, evaluate. Good for brief-driven design runs that must use what was found.
• Simple retry - Do the task, evaluate, retry with fixes. Minimal loop when you want a few honest retries without a long procedure stack.

Create your own

Create loop lets you stack procedures: start, generate, evaluate, improve, research, plan, act, clarify, finish. Name the loop, set max iterations, and write a rubric that says what a top score looks like.

Choose a Run mode: Full cycle runs every procedure in each pass (best for writing and design briefs). One procedure steps through them one pass at a time (best for tool-heavy work, so act phases are not crowded).

Under Advanced you can tune when the loop gives up on a plateau (stagnation stop), turn the independent judge on or off, or sample multiple candidates per generate pass and let the next pass pick the strongest.

Scores you can trust

Evaluation runs as an independent judge - a fresh look at the work, not the agent grading its own answer. Give each loop a rubric describing what a 10 looks like, so scores are anchored to your standard rather than general impressions.

If quality dips on a later pass, the block still returns the best-scored version. When passes stop improving, the loop stops early instead of burning remaining passes on repeats.

When studios reach for loops

• One draft is never enough - competition narratives, zoning memos, and concept write-ups need a few honest refine passes.
• You have a clear bar for quality and want the agent to chase it instead of you rewriting the prompt three times.
• You want critique without adding an Evaluator block and wiring a critic graph for every Agent.
• Tool-heavy architecture work needs plan → act → evaluate → improve with room for tools on each act pass.
• Research should visibly show up in the output - Research then generate stops rewarding answers that ignore the context.

How to start

• Open an Agent block and find the Loop field.
• Choose Add loop for a built-in, or Create loop to stack your own procedures.
• On the selected loop chip, edit the max number next to the name to cap passes (1-20).
• Add a Rubric that describes what a top score looks like.
• Run. Open the block output to see the score trajectory and which pass won.
• Remove the loop from the chip anytime - the Agent goes back to a single pass.

Available now

Agent block loops are available in the workspace today - built-ins and custom procedure stacks, independent judge, best-result output, and early stop on plateau. They sit next to skills, manuals, and tools on the same Agent block.

Design agents are how the studio does the job. Loops are how one Agent keeps improving until the work meets the bar you set.

Design Agents · VERSUR Agent · VERSUR Auto · Changelog